Teach17

AI tutoring is no longer a distant possibility. The UK government is already funding the development and testing of AI tutors designed to support pupils in English, mathematics, science and modern foreign languages.

The attraction is understandable. An AI tutor could provide an explanation, hint or practice question at the moment a pupil becomes stuck. It might give more young people access to the kind of individual support currently available mainly to families who can afford private tuition.

However, there is an important difference between helping a pupil learn and helping them finish their work. Schools will need to make sure that AI tutors preserve thinking, effort and independence rather than removing them.

Here are 17 things schools need to know.

1. An AI tutor should teach, not simply answer

A general chatbot usually tries to provide the user with a quick and complete response. That is not necessarily what a pupil needs.

A genuine AI tutor should diagnose the difficulty, ask a useful question, provide a small hint and give the pupil another opportunity to think.

For example, when a pupil cannot rearrange an equation, the AI should not immediately display the complete solution. It might first ask which operation is being applied to the unknown quantity and what inverse operation could be used.

The quality of an AI tutor should be judged by the thinking it encourages, not the speed at which it produces answers.

2. Government-backed AI tutoring trials are beginning in England

In January 2026, the government announced plans to develop AI tutoring tools that could eventually support up to 450,000 disadvantaged pupils each year.

The AI Tutoring Tools Pioneer Programme is focusing initially on pupils in Years 9 and 10. The proposed tools will cover English, mathematics, science and modern foreign languages.

Up to eight companies were invited to join the programme, with each successful organisation eligible to receive £300,000. The tools are being designed with teachers and are due to be tested in schools under teacher supervision. Successful products could become available nationally from 2027.

This means schools are likely to hear much more about AI tutoring during the next two years.

3. The plans are partly an attempt to reduce inequality

High-quality one-to-one tuition can produce approximately five additional months of progress. It appears to work best when sessions are regular, targeted and explicitly linked to classroom teaching.

Unfortunately, access to private tutoring remains unequal. The government hopes AI will make individual support available to pupils whose families could not otherwise afford it.

That is an ambitious and worthwhile aim. However, giving pupils access to software will not automatically reproduce the expertise, encouragement and accountability of a good human tutor.

4. There is some encouraging evidence

A 2025 study involving university physics students found that a carefully designed AI tutor produced greater learning gains than an active-learning lesson covering the same material.

The tutor followed a structured sequence, used established teaching principles and required pupils to engage with the problems. It was not simply an unrestricted chatbot.

An exploratory UK trial also involved 165 pupils across five secondary schools using an AI-supported mathematics tutor. Pupils receiving AI-supported tutoring performed at least as well as those receiving human tutoring on the outcomes measured.

However, the AI’s messages were supervised by expert tutors, and the study was a preprint rather than a completed peer-reviewed evaluation. It is promising evidence, but it does not prove that unsupervised AI tutoring will work equally well.

5. The wider evidence is still developing

A systematic review examined 28 studies of intelligent tutoring systems involving 4,597 school pupils. It found generally positive effects on learning and performance.

However, the advantage was smaller when AI tutoring was compared with other well-designed digital tutoring systems. Many studies were also relatively short and involved limited numbers of pupils.

The evidence therefore suggests potential rather than certainty. Schools should be cautious about claims that a particular product has already “transformed learning”.

6. Unrestricted AI can make pupils appear more successful

One of the most important studies in this area examined pupils using generative AI for mathematics.

Access to a standard AI assistant improved performance while pupils were using it. However, when the AI was removed, some pupils performed worse than pupils who had never used it.

A more carefully designed version, which provided hints and educational guardrails, reduced this problem.

This distinction matters. Completing more questions during an AI session does not necessarily mean that more learning has taken place.

7. AI tutors must protect productive struggle

Learning often requires pupils to experience uncertainty, make mistakes and persevere.

An AI tutor that intervenes too quickly may remove the exact thinking a pupil needs to practise. A pupil should not receive a hint simply because they have paused for five seconds.

Schools should look for tools that allow teachers to control the amount of support provided. Useful settings might include:

• hints only after an attempted answer
• a limit on the number of hints
• no complete solution until the task is finished
• a requirement for pupils to explain their reasoning
• independent questions after the supported practice

The best tutor does not prevent struggle. It makes that struggle manageable and purposeful.

8. Good AI tutoring should use gradual hints

Effective tutoring usually provides the smallest amount of help needed to move learning forward.

A useful sequence might be:

1. Ask the pupil to explain what they understand.
2. Draw attention to the relevant information.
3. Remind them of an appropriate strategy.
4. provide part of a worked example.
5. Show the complete method only when necessary.
6. Give the pupil a similar problem to complete independently.

This approach is much more educationally useful than repeatedly producing polished explanations that pupils can read without thinking.

9. Immediate feedback could be a genuine advantage

A teacher cannot provide detailed individual feedback to every pupil at the exact moment they become stuck.

An AI tutor might immediately notice that a pupil has:

• selected the wrong formula
• confused mass and weight
• forgotten to convert a unit
• misinterpreted a command word
• used the correct method but made an arithmetic error

The pupil could then correct the error while the reasoning is still fresh.

This is one of the strongest potential benefits of AI tutoring, particularly during independent practice. It will only work, however, when the feedback is accurate and directly related to the curriculum.

10. AI tutoring should develop metacognition

A good tutor should help pupils become more aware of how they learn.

Instead of always giving information, an AI tutor could ask:

What is the question asking you to find?

What knowledge might be useful here?

Where did your method begin to go wrong?

How could you check your answer?

What would you do differently next time?

Metacognition and self-regulation have a strong evidence base. They involve pupils planning, monitoring and evaluating their own learning.

An AI tutor should support these habits rather than becoming a permanent substitute for them.

11. Teachers must remain in control

An AI tutor does not know a pupil in the same way as their teacher.

It may not recognise that the pupil is anxious, rushing, guessing, disengaged or deliberately entering random answers. It cannot fully understand the classroom context or the relationships surrounding the pupil.

Teachers should decide:

• which pupils use the tool
• which topics it supports
• when it can be used
• how much help it provides
• what information is shared with staff
• when a pupil needs human intervention

The government has stated that its proposed AI tutors should complement face-to-face teaching rather than replace it.

That principle must survive beyond the trial stage.

12. Schools need evidence of independent learning

AI platforms can collect impressive amounts of data. They may report time spent, questions completed, hints requested and topics practised.

These figures are useful, but they do not prove that learning has transferred.

Schools should check what pupils can do after the AI has been removed. This might involve:

• a short exit question
• an unaided quiz the following lesson
• a delayed retrieval task
• an unfamiliar application question
• asking pupils to explain the method aloud

The central measure should be what the pupil can now do independently.

13. AI tutors will sometimes be wrong

Generative AI can produce incorrect information in confident and convincing language.

In mathematics and science, it might make an arithmetic error, apply an inappropriate formula or confuse two related concepts. In English or humanities subjects, it may invent evidence, misrepresent a text or produce a quotation that does not exist.

A pupil with weak prior knowledge may be the least able to recognise these errors.

Schools should test an AI tutor with real curriculum questions before using it. Teachers should also be able to report inaccurate responses and see how quickly the provider corrects recurring problems.

14. Safeguarding must be built into the product

Pupils may disclose personal information to a chatbot or begin treating it as a trusted companion.

The Department for Education’s updated safety standards state that educational AI should avoid manipulative design, emotional simulation and language that encourages secrecy or isolation. Products should also include appropriate safeguarding and mental-health protocols.

The government’s AI tutoring programme also requires products to be age-appropriate, safe and aligned with the National Curriculum.

Before adopting a tool, schools need clear answers about what happens when a pupil enters content relating to abuse, self-harm, bullying or another safeguarding concern.

15. Pupil data needs careful protection

AI tutoring systems may collect considerably more information than an ordinary worksheet.

A conversation could reveal a pupil’s attainment, misconceptions, reading ability, confidence, SEND needs or emotional state.

Schools should establish:

• what data is collected
• where it is stored
• who can access it
• how long it is retained
• whether it is used to train AI models
• whether parents and pupils have been properly informed

The Information Commissioner’s Office advises educational institutions to consider data-protection obligations from the beginning, particularly when children’s information is involved.

The government has said that identifiable pupil data will not be made public through its tutoring programme and that pupil work will not be used to train AI without parental permission.

16. Technology can widen the disadvantage gap

AI tutoring is being promoted partly as a way to improve outcomes for disadvantaged pupils. That does not guarantee that it will do so.

Pupils may need a suitable device, reliable internet connection, headphones, a quiet place to work and confidence using the platform. Some will require adult encouragement to remain engaged.

An EEF review warned that educational technology can widen the disadvantage gap when access, implementation and pupil needs are not properly considered.

Schools should examine participation and outcomes for different pupil groups rather than reporting only an overall average.

17. Schools should begin with a focused trial

AI tutoring should not be introduced across every subject and year group at once.

A stronger starting point would be a clearly defined problem, such as:

• helping Year 9 pupils practise manipulating equations
• supporting Year 10 vocabulary retrieval in French
• giving feedback on short GCSE science calculations
• helping pupils identify gaps before a mathematics assessment
• providing structured reading-comprehension practice

The school should agree success criteria before the trial begins. These might include improved unaided assessment scores, fewer repeated misconceptions, increased homework completion or positive pupil and teacher feedback.

The government is also investing an additional £23 million to expand its EdTech Testbeds programme, allowing more than 1,000 schools and colleges to evaluate AI and assistive technologies. National benchmarks are being developed to assess whether educational AI is accurate, safe, age-appropriate and curriculum-aligned.

Schools should use the same evidence-led approach locally. Start small, compare outcomes and expand only when the technology is genuinely improving learning.

AI tutoring could become a valuable addition to schools. It may provide immediate feedback, extra practice and individual explanations on a scale that would previously have been impossible.

However, the most successful AI tutor will not be the one that completes the greatest number of tasks or produces the most impressive dashboard. It will be the one that asks useful questions, protects productive struggle and helps pupils become less dependent on support over time.

The goal should not be for every pupil to have an AI that can answer anything. It should be for every pupil to receive the right help, at the right moment, while still doing the thinking for themselves.